When people think about building AI, their first thought is usually big tech or the billion-dollar AI startups like OpenAI or Anthropic. But in India, the approach is a little different. Here, the government is leading in providing funding and setting up the required infrastructure and platform for the entire population.
Over the last three years, government-led Bhashini has quietly built language models, translation platforms, and tools that now power 10 million translation requests daily, with the Bhashini app crossing a million downloads.
Yet, that’s just the beginning. India’s flagship language AI mission isn’t trying to outpace OpenAI or the growing wave of Indian AI startups. It’s laying the foundation for them.
“Think of Bhashini like NPCI,” Amitabh Nag, CEO of Bhashini and director of IndiaAI, told AIM on the What’s the Point podcast. “We’re not trying to become the biggest company in AI—we’re building public infrastructure that makes such growth possible,” he said.
But this isn’t just about scale. It's about access.
As OpenAI and other Western players eye India, there’s growing curiosity about whether local initiatives like Bhashini and homegrown AI startups can compete. Nag dismissed this concern. “It’s not a clash. If you’re solving a problem through a strong use case, you’ll survive OpenAI, you’ll survive anyone,” he said.
“India’s ChatGPT Moment” Isn’t a Product
Bhashini's most ambitious bet is speech-to-speech translation. “The idea is to have a push-button system that makes you multilingual,” Nag said. “Automatic language detection, speaker ID, diarization—it’s all part of the pipeline.” And in some benchmarks, Bhashini’s speech systems already outperform OpenAI in latency.
“We were the first to talk about 100% accuracy and voice-first systems when the industry was still unsure,” he said. “We’re not far. But it’s not about building a replica. We had to start with creating the digital data itself in Indian languages—something no one else in the world had done.”
“Bhashini is useful when I want to talk to a taxi driver in Kannada, while I don’t speak the language. It works, and it works affordably,” he said. “ChatGPT can translate, sure—but that’s not the point. We’re building something for Bharat.”
Bhashini’s mission is rooted in public good, not market dominance. “If someone else can do it better, that’s fine. The idea is to create healthy competition and show what’s possible.”
That’s what led to initiatives like Bhasha Daan, where citizens contributed parallel corpora through mobile apps to help train models in low-resource languages such as Bodo, Manipuri, Khasi, Mizo, and even scripts like Kaithi, which are still used in Bihar’s land records.
The reason Bhashini exists is precisely because private players wouldn’t start here. “Try convincing a startup to build for Bodo,” Nag said. “That’s why the government stepped in.”
Not India’s Hugging Face, Something Deeper
One of the ambitious projects Nag is leading is IndiaAI Mission’s AI Kosh—Bhashini’s model and dataset repository. While this seems like India’s version of Hugging Face, Nag doesn’t think so. “We’re building something deeper. Think of it as the currency of AI for India.”
AI Kosh will not only house datasets and models, but also enable entire use-case pipelines, acting as a KoshaGraph for future AI systems. “The Internet in India didn’t become just another advertising engine—it became a tool for inclusion. AI will follow that path too.”
Bhashini is not alone. It’s working alongside startups, research labs, and platforms like AI4Bharat. “Some are building governance frameworks, others foundation models. It’s all complementary. Eventually, the jigsaw will fit.”
The mission’s next frontier for Bhashini is Tier 2 and Tier 3 cities, especially in states like Madhya Pradesh, Uttar Pradesh, Jharkhand, Bihar, and the Northeast. “We’re approaching state governments with MoUs, not as a top-down directive, but as a co-creation exercise,” Nag explained. “The idea is for states to adopt, fine-tune, and own the models over time.”
Sovereignty Isn’t Just About Data
India emphasises building AI domestically while Western countries don’t show the same urgency. Nag drew an analogy to child-rearing to clarify why that is the case.
“AI learns like a child—it absorbs values, culture, and behaviour from its environment. If we want our models to respect our linguistic nuances, social behaviours, and cultural codes, we must build them ourselves,” he explained. For instance, Nag noted that a child from Telangana addresses his or her elders differently than a person from Tamil Nadu regarding the language.
“It’s about encoding the fundamentals of our society into AI systems.”
This is also why the Bhashini app remains in beta—but that’s deliberate. “It works well in 35+ languages; we added another just yesterday,” he said. The beta tag is less a sign of immaturity and more a signal of continuous improvement. “Beta just means we’re on the road to improvement—it’s not traditional beta,” Nag clarified.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.